Machine learning in investor support tools

Abstract

With bank deposit interest rates plummeting to historic levels and even reaching negative values in certain cases, investing in the stock markets is becoming more attractive to the general community. This research aims to evaluate the performance of machine learning technology in stock trading, by aiming to learn the relationships between stocks, news events and time. The proposed technology uses hybrid machine learning component, based on ANN (Artificial Neural Network) and K-Means Classifiers; enabling a one pass training process that could be executed every time a prediction is generated, thus making sure that the prediction is updated as possible. Parameters such as the stock, training data interval and prediction period could be adjusted as necessary on a per-prediction basis. A prototype implementation of the proposed technology has been created in C# as a Windows Forms application. Import routines and web crawlers for stock prices have been used to retrieve stock price and news data. The prototype has been evaluated scientifically using quantitative methods, on a selection of local Maltese stocks as well as foreign ones; and has performed favourably in each case, with the Maltese stock predictions achieving a better margin of error than their foreign counterparts. Prediction error margins ranged from 0.64% to 3.63% in single day predictions using 3 months of training data. Worst case predictions of 7 trading days in the future resulted in error margins of 1.42% to 13.32%, again with 3 months of training data. These results confirm that machine learning technology could be applied to stock price prediction based on news data, and this research could provide a solid base for further research in the area.